Re: [PATCH] net/mlx5e: fix high stack usage

[Saeed Mahameed <saeedm@mellanox.com>]

On Fri, 2018-11-02 at 16:15 -0600, Jason Gunthorpe wrote:
> On Fri, Nov 02, 2018 at 10:07:02PM +0000, Saeed Mahameed wrote:
> > On Fri, 2018-11-02 at 14:39 -0700, Eric Dumazet wrote:
> > > 
> > > On 11/02/2018 02:05 PM, Saeed Mahameed wrote:
> > >  
> > > > temp will be mem copied to priv->stats.sw at the end,
> > > > memcpy(&priv->stats.sw, &s, sizeof(s)); 
> > > > 
> > > > one other way to solve this as suggested by Andrew, is to get
> > > > rid
> > > > of
> > > > the temp var and make it point directly to priv->stats.sw 
> > > > 
> > > 
> > > What about concurrency ?
> > > 
> > > This temp variable is there to make sure concurrent readers of
> > > stats
> > > might
> > > not see mangle data (because another 'reader' just did a memset()
> > > and
> > > is doing the folding)
> > > 
> > > 
> > > mlx5e_get_stats() can definitely be run at the same time by
> > > multiple
> > > threads.
> > > 
> > 
> > hmm, you are right, i was thinking that mlx5e_get_Stats will
> > trigger a
> > work to update stats and grab the state_lock, but for sw stats this
> > is
> > not the case it is done in place.
> 
> That was my guess when I saw this.. the confusing bit is why is there
> s and temp, why not just s?
> 

silly cut and paste from mlx4 !

> > BTW memcpy itself is not thread safe.
> 
> At least on 64 bit memcpy will do > 8 byte stores when copying so on
> most architectures it will cause individual new or old u64 to be
> returned and not a mess..
> 
> 32 bit will always make a mess.
> 
> If the stats don't update that often then kmalloc'ing a new buffer
> and
> RCU'ing it into view might be a reasonable alternative to this?
> 

kmalloc is not an option, stats update too often, priv->stats.sw is
already a temp, it is needed only to accumulate channels stats into it
and then report them to ethtool buffer or ndo_get_stats , both will
copy the priv->stats.sw to their own buffers, so there can only be two
use cases (two threads), maybe it is best if we maintain two priv-
>stats.sw one for each case and use it as a temp for each thread.

> Jason